Comparison of Ridge Regression and Partial Least Squares Regression for Estimating Above-Ground Biomass with Landsat Images and Terrain Data in Mu Us Sandy Land, China

Abstract

Although there has been no consensus on the proper way to assess the desertification status of a piece of land, plant biomass estimation is an undeniably important part of monitoring and assessing land desertification. Considering the limitations of the single vegetation index, the ridge regression and partial least squares (PLS) regression approaches were used to construct multivariate models and then estimate above-ground biomass in a semi-arid area, Mu Us sandy land, in China. The predictive performances of these approaches were compared using a test data set (n = 20), while the models were constructed using a calibration data set (n = 50). Results showed that when the relative root mean square error (RMSE) was about 61%, a lower error was produced by the reduced PLS regression model, while the reduced ridge regression model was 64%. The full ridge regression model performed poorly when the RMSE % was only 87%, and the predicted biomasses were not correlated to the actual ones (r = 0.0181, n = 20). This study suggests that PLS regression based on Landsat ETM+ data provides a better approach for the above-ground biomass estimation, and that variable selection can contribute to better estimation.